Discrepancy in mAP Results between Provided Model and Self-Trained Model on YTVIS2019 Validation Set

[SkyKingL]

Hello,

I've been experimenting with your VIS model and I've encountered an issue that I hope you can help me clarify.

I trained the model on a server equipped with eight V100-32GB GPUs using the mfvis_nococo/scripts/train_8gpu_mask2former_r50_video.sh script, without any modifications to the parameters. The test results I obtained on the YTVIS2019 validation set are as follows:

For comparison, I also tested the trained weight you provided on the same YTVIS2019 validation set and obtained these results:

As you can see, there is a minor difference in the mAP scores. I was unable to achieve the mAP=42.5 as reported in your documentation. Could you please provide some insight as to why this discrepancy might be occurring?

I appreciate your time and look forward to your response. Thank you.